Close to half of office workers say they are uncomfortable telling their manager they used AI.
Not uncomfortable using it. Uncomfortable admitting it.
If you are trying to work out why nothing much is happening in your department, start there rather than with the training budget. The blocker is not capability, and it is not access. It is that the people you are asking to try something have quietly decided it is safer not to mention it.
Where the number comes from
Marcus Loh, a communications professional writing in TNGlobal Insider on 21 August, cites a Slack survey with that finding.
Worth being precise about what that is. It is an opinion column, not reporting and not research, and the survey is cited without a link, so treat the figure as directional rather than as established fact.
Even at its most cautious, the shape holds. A large minority of people are using these tools and choosing not to say so, and Loh’s argument is that what they are avoiding is a judgement about them rather than a judgement about the work.
What Loh was actually arguing
His subject is Anthropic’s text watermark, which now embeds an invisible signature in Claude’s output to comply with the European AI Act’s transparency requirements.
His concern is not the technology. It is that a detectable marker turns a question of quality into a question of provenance:
“A watermark can indicate that Claude was involved somewhere in the process, but it cannot measure the extent of that involvement.”
That is the crux. A marker tells you a tool was present. It tells you nothing about whether the person understood the output, checked it, disagreed with parts of it, or rewrote half of it. So the same signal covers the person who used AI well and the person who pasted without reading, and if the organisation has no way to tell those apart, everyone has an incentive to stay quiet.
I think Loh is right that the workplace consequence of better detection is stigma rather than transparency. But the more useful reading for a manager is one step earlier, because the stigma is already there and no watermark created it.
Why this is a management problem
Consider what a department looks like where half the AI use is undeclared.
You cannot see what is working. Somebody has found a way to cut two hours off a weekly report and nobody else will ever learn it, because sharing it means admitting to it.
You cannot fix what is not working. The person quietly pasting unchecked output into client documents is doing the single most dangerous version of this, and they are the least likely of anyone to raise their hand.
You cannot train people in a practice nobody admits to. Every training conversation starts by asking people to describe what they currently do, and you will get a room of blank faces.
And you cannot manage the risk, because the risk has moved somewhere you cannot look at it. That is worse than a team not using AI at all. A team not using it has a capability gap. A team using it in secret has a capability gap and a governance gap, and only one of those shows up in a report.
The standard that actually works
The fix is not better detection. It is a stated standard that makes disclosure safe, and the version that works is about ownership of the finished work rather than disclosure of the tool.
Sophie Alpert put it about as well as it can be put in her own written policy on AI-assisted writing:
“You must stand behind every idea and every sentence in your docs. It is your responsibility to make sure that the entire document is representative of your own thoughts before you share it.”
Her corollary is the useful half: it is not acceptable to reply “oh sorry, AI wrote that”.
Notice what that standard does. It stops asking whether a tool was involved, which is unanswerable and unenforceable and a bit undignified. It asks whether you are prepared to defend what left your desk, which is exactly what you already ask of everyone about everything.
It also removes the cheating frame entirely. Nobody thinks you cheated by using a spreadsheet, a calculator or a colleague, because in each case you remain answerable for the result. The uncomfortable half of your team is not confused about the ethics. They are waiting to find out which rule they are being judged against, and in the absence of a stated one, the safe assumption is the harshest one.
The ten-minute version
You do not need a policy document. You need to say something out loud, and then write it down somewhere findable.
Three sentences will cover it. Using AI to help produce your work is fine and expected. You are responsible for what you send, the same as always, whatever helped you produce it. If it goes wrong, “the AI wrote it” is not an answer, and neither is anything else you would not have accepted before.
Then say the second part, which matters more: nobody is in trouble for anything they did before this conversation. Without that, you have announced a standard and simultaneously created a group of people with a reason to keep hiding.
Say it in a team meeting rather than an email, so people can see the reaction on your face while you say it. Then go first: describe something you used it for this week, including a bit where it was wrong and you caught it. The point of going first is not modelling good practice. It is proving that admitting to it is survivable.
The question
Has anyone in your organisation said out loud whether using AI is something you own or something you hide?
Most have not. Most have a policy about data and nothing at all about disclosure, which leaves every individual to guess. The guess they make is the cautious one, and the cost of that caution is the thing you were blaming on the training budget.
Sources: Marcus Loh, “Could Anthropic’s watermark turn AI assistance into a mark of suspicion?”, TNGlobal Insider, 21 August 2026, an opinion column. Sophie Alpert on acceptable use of AI in writing.